CN101714255A - Abnormal behavior detection device - Google Patents

Abstract

The present invention provides an abnormal behavior detection device, which can calculate the degree of abnormal behavior of people and animals in an image, and accurately determine whether abnormal behavior has occurred according to the degree of abnormality. In the abnormal behavior detection device according to the present invention, the degree of abnormality of the video of the monitoring object acquired by the video acquisition unit is calculated, and whether abnormal behavior has occurred is determined from the calculated degree of abnormality based on the threshold value. Since it is possible to accurately determine whether an abnormality has occurred based on the degree of abnormality in the image, when an abnormality occurs, it is possible to promptly take measures against the abnormality by issuing a warning or notifying security personnel.

Description

Abnormal Behavior Detection Device

This application is a divisional application, the application number of its parent application is: 200710004061.5, the application date: 2007.1.23, the name of the invention: abnormal behavior detection device

technical field

The present invention relates to an abnormal behavior detection device for detecting abnormal behaviors of humans, animals and the like.

Background technique

In response to social instability such as an increase in the incidence of crimes, the number of cameras installed to monitor suspicious persons and vehicles is increasing. When using so many cameras for surveillance, in order to effectively monitor the surveillance area with limited surveillance personnel resources, surveillance assistance technology is required.

As such monitoring assistance technology, for example, Japanese Patent Application Laid-Open No. 2005-92346 of Patent Document 1 discloses a "method and device for extracting feature quantities from three-dimensional data". It discloses a method of extracting feature quantities of moving images called stereoscopic high-order local autocorrelation features. Furthermore, a method of applying this feature quantity to behavior recognition and walking posture authentication has also been disclosed.

In addition, Non-Patent Document 1 discloses a method of calculating the degree of abnormality of a person's behavior in an image using stereoscopic high-order local autocorrelation features.

Patent Document 1: Japanese Patent Laid-Open No. 2005-92346

Non-Patent Document 1: "Detecting Abnormal Actions from Moving Images of Multiple People", Takuya Minamisato, Shinyuki Otsu, Research Report 2004-CVIM-145 of the Society for Information Processing, September 11, 2004

The prior art described above calculates the degree of abnormality in the behavior of people, animals, etc. in the video as a scalar quantity, and there is a problem that it cannot be judged immediately whether abnormal behavior has actually occurred.

Contents of the invention

The present invention is made based on the above problems, and its object is to provide an abnormal behavior detection device that can calculate the degree of abnormality in the behavior of people, animals, etc. abnormal behavior.

In the abnormal behavior detection device according to the present invention, there are: an image acquisition unit that acquires an image of a monitoring object; an abnormality degree calculation unit that calculates an abnormality degree of the image acquired by the image acquisition unit; and an abnormality judgment. A unit that judges whether an abnormal behavior has occurred from the abnormality degree calculated by the abnormality degree calculation unit based on a threshold value, and is characterized in that it also has an error indicating the relationship between the false alarm rate and the non-report rate and the threshold value. The graph is displayed on the display section of the screen.

According to the present invention, it is possible to accurately determine whether an abnormal behavior has occurred based on the degree of abnormality in the video. Therefore, when an abnormality occurs, it is possible to promptly take corresponding measures against the abnormal behavior by issuing a warning or notifying the security personnel.

Description of drawings

FIG. 1 is a block diagram showing the functional structure of an abnormal behavior detection device as an embodiment of the present invention.

FIG. 2 is a flowchart showing a flow of processing at the time of abnormality judgment.

FIG. 3 is a block diagram showing a functional configuration of an abnormality degree calculation unit.

FIG. 4 is a flowchart showing the flow of abnormality level calculation processing.

FIG. 5 is an explanatory diagram showing frames used when calculating stereo high-order local autocorrelation.

Fig. 6 is an explanatory diagram of a lattice structure (mask pattern) of stereo high-order local autocorrelation.

FIG. 7 is a flowchart showing the flow of calculation processing of stereo high-order local autocorrelation features.

FIG. 8 is a flowchart showing the flow of calculation processing of a transformation matrix.

FIG. 9 is an explanatory diagram of calculation processing of a local space.

FIG. 10 is an explanatory diagram of a method of evaluating the degree of abnormality.

FIG. 11 is a flowchart showing a flow of processing at the time of determination threshold value calculation.

FIG. 12 is an explanatory diagram of the maximum degree of abnormality in each scene.

Fig. 13 is an explanatory diagram of false alarm rate and non-report rate.

Fig. 14 is an explanatory diagram of error curves.

FIG. 15 is an explanatory diagram of local space determination processing.

FIG. 16 is an explanatory diagram of an example of a monitor screen.

FIG. 17 is a flowchart showing the flow of abnormality determination processing at three levels.

Fig. 18 shows an elevator apparatus provided with an abnormal behavior detection device according to the present invention.

In the figure: 10-abnormal behavior detection device, 20-elevator control device, 30-camera, 40-elevator car, 50-tail cable, 100-image acquisition department, 102-abnormal degree calculation department, 104-abnormality judgment department, 106-judgment threshold, 108-judgment result, 110-judgment threshold calculation unit, 112-notification unit.

Detailed ways

Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing the functional structure of an abnormal behavior detection device as an embodiment of the present invention. This device is composed of an image acquisition unit 100 , an abnormality degree calculation unit 102 , an abnormality determination unit 104 , a judgment threshold value calculation unit 110 , and a notification unit 112 , and detects abnormal behaviors based on images of monitoring objects acquired by the image acquisition unit 100 . The description will be given below in order.

The video acquisition unit 100 is an imaging device such as a video camera or a video playback device such as a video recorder, and is used to obtain video to be input to the device. The imaging device is used when the live image being captured is used as an input. The video playback device is used when receiving video accumulated in the past as input.

The abnormality degree calculation unit 102 is used for calculating the abnormality degree of the images acquired by the image acquisition unit 100 . Wherein, the so-called degree of abnormality is a scalar quantity, which represents the degree of abnormal behavior of moving objects such as people and animals in the video.

The abnormality determination unit 104 determines whether abnormal behavior has occurred based on the abnormality level calculated by the abnormality level calculation unit 102 , and outputs the result as a determination result 108 . Using the judgment threshold 106 as a judgment criterion, when the degree of abnormality is smaller than the judgment threshold 106 , it is judged that no abnormal behavior has occurred. On the contrary, when the degree of abnormality is equal to or greater than the judgment threshold 106, it is judged that abnormal behavior has occurred.

The determination threshold calculation unit 110 is used to calculate the determination threshold 106 required when the abnormality determination unit 104 performs determination processing. Among them, the determination threshold calculation unit 110 performs calculations based on the determination result 108 so that the determination accuracy of the abnormality determination unit 104 becomes the optimum determination accuracy.

The notification unit 112 notifies an external device that an abnormal behavior has occurred when an abnormal behavior has occurred based on the determination result 108 . The external device that has received the notification can output the alarm in the form of voice, and can also output the alarm to the monitoring screen. Moreover, it is also possible to stop the operation of devices such as elevators based on safety considerations. In addition, it is also possible to notify the monitoring center and mobile terminals, etc. by means of long-distance communication, so as to prompt them to take measures.

The abnormality degree calculation unit 102 , the abnormality determination unit 104 , the determination threshold value calculation unit 110 , and the notification unit 112 can be realized by a CPU or an arithmetic processing device such as a CPU, or a personal computer. In addition, the judgment threshold and judgment results are stored in a storage device such as a semiconductor memory, and can be read out at any time and used in various calculations.

Referring to the flow chart of FIG. 2 , the processing flow when the abnormal behavior detection device of this embodiment is used for abnormal judgment will be described below.

In step 200, the processing from step 202 to step 210 is repeated at a predetermined frequency set in advance until the user issues an end instruction.

In step 202 , the video acquired by the video acquisition unit 100 is read as digital data by the abnormality degree calculation unit 102 .

In step 204 , the abnormality degree of the image acquired in step 202 is calculated by the abnormality degree calculation unit 102 .

In step 206, it is judged by the abnormality determination unit 104 whether abnormal behavior has occurred or not by using the degree of abnormality calculated in step 204.

In step 208, the judgment result of step 206 is evaluated, and when it is judged that an abnormal behavior has occurred, step 210 is executed.

In step 210 , the notification unit 112 notifies the external device that an abnormal behavior has occurred.

The internal configuration of the abnormality degree calculation unit 102 in FIG. 1 will be described in detail below with reference to the block diagram of FIG. 3 . As described above, the abnormality degree calculation unit 102 calculates the abnormality degree of the image acquired by the image acquisition unit 100 as a scalar quantity, and outputs it to the abnormality determination unit 104 . The abnormality degree calculation unit 102 is composed of an activity extraction unit 300 , a feature quantity calculation unit 302 , a feature quantity conversion unit 304 , and an abnormality degree evaluation unit 308 . The following will explain in order.

The motion extraction unit 300 extracts a moving part from the video acquired by the video acquisition unit 100 . Its purpose is to remove static parts such as the background that are irrelevant to the judgment of abnormal behavior. A known video processing method (see Japanese Patent Application Laid-Open No. 2005-92346, etc.) can be used to extract the moving portion. For example, a method of acquiring only the difference between two frames, or a method of acquiring the difference between frames after performing edge extraction processing, etc. may be used. In addition, in order to remove the influence of interference such as illumination fluctuations, after obtaining the difference between frames, the pixel value can be set to 0 or 1, and additional binarization processing can be performed.

The feature quantity calculation unit 302 calculates the feature quantity of the video generated by the motion extraction unit 300 . For calculation, a known stereoscopic high-order local autocorrelation feature is used (for example, refer to Japanese Patent Application Laid-Open No. 2005-92346). In this method, geometric features of voxel data consisting of three consecutive frames of video are calculated as 251-dimensional feature vectors. The calculation method of this feature quantity will be described later. In addition, when calculating the feature quantity of an image, an optical flow calculation method may also be used. The so-called optical flow calculation method is a method that focuses on the small area of the image and calculates the motion between frames as a vector. discussion. All the components of the vectors calculated by the optical flow calculation method may be combined as feature quantities. In addition, statistical quantities such as average and dispersion of vectors may be used as feature quantities.

The feature quantity conversion unit 304 performs linear conversion on the feature quantity vector calculated by the feature quantity calculation unit 302 using the transformation matrix 306 . Through this conversion, the component of the abnormal behavior contained in the feature quantity vector is extracted. Wherein, assuming that the feature vector calculated by the feature calculating section 302 is x, the conversion matrix 306 is A, and the converted feature vector is x', the conversion can be expressed by the following formula (1).

x’＝Ax (1)

The transformation matrix 306 is a matrix obtained by multivariate analysis such as principal component analysis, and its calculation method will be described later. When the 251-dimensional high-order local autocorrelation feature is used as the feature quantity of the image, the size of the transformation matrix 306 is (252-n)×251 (n=1, 2, . . . , 251). And, the feature quantity linearly transformed by this matrix becomes a 252-n-dimensional vector.

The degree of abnormality evaluation section 308 calculates the degree of abnormality by evaluating the degree of deviation between the new feature amount vector calculated in the feature amount conversion section 304 and the normal data 310 . And, the calculation result is output to the abnormality determination unit 104 . Among them, the normal data 310 is a collection of feature quantities of normal behavior. The specific calculation method of the degree of abnormality will be described later.

Hereinafter, referring to the flowchart of FIG. 4 , the abnormality degree calculation process in step 204 of FIG. 2 will be described in detail.

In step 400 , the motion extracting unit 300 extracts a moving part from the video acquired by the video acquiring unit 100 .

In step 402 , the feature amount of the video generated in step 400 is calculated by the feature amount calculation unit 302 .

In step 404, the feature vector calculated in step 402 is linearly transformed by the feature transform unit 304 to generate a new feature vector.

In step 406 , the degree of deviation between the new feature vector calculated in step 404 and the normal data 310 is evaluated by the abnormality degree evaluation unit 308 to calculate the degree of abnormality.

Hereinafter, referring to FIGS. 5 to 7 , the moving image feature amount calculation process described in step 402 of FIG. 4 will be described in detail.

FIG. 5 is an explanatory diagram of input data of the above-mentioned stereo high-order local autocorrelation feature. The calculation object of the feature quantity is a moving image, that is, consecutive frames (images) in time series. In order to calculate stereo high-order local autocorrelation features, at least three frames are required. For example, when a frame 500 with a frame number n is given, the frame, the frame 502 and the frame 504 (corresponding to frame numbers n-1 and n-2, respectively) before this frame become feature quantities Computed object.

Assuming that the resolution of the frame is h pixels in the vertical direction and w pixels in the horizontal direction, by combining three frames, a voxel (cube) of h×w×3 can be formed. In the three-dimensional high-order local autocorrelation feature calculation method, all elements of the voxel are moved sequentially using a 3×3×3 lattice structure 506 to extract features.

In addition, in the present embodiment, the case where three consecutive frames are set as the processing target has been described, but any f number of frames may be set as the processing target. At this time, with voxels of h×w×f as processing objects, the average feature value of f frames of moving images is calculated.

FIG. 6 is an illustration of a lattice structure used when calculating stereo high-order local autocorrelation features. The lattice structure is used to calculate the local relevant features of voxels, which consists of 3×3×3 voxels.

Mode 1 is a counting mode that counts the number of voxel 600 located at the center where the pixel is 1 when sequentially scanning the voxel data of the input image. Similarly, mode 2 is a mode for counting the number of times when the voxel 602 is 1 in addition to the central voxel 604 .

There are 251 lattice structures in the stereoscopic high-order local autocorrelation feature of the binary image, and by counting the number that satisfies each mode, the features of the input image can be extracted as a 251-dimensional feature vector.

Referring to the flowchart of FIG. 7 , the moving image feature amount calculation process described in step 402 of FIG. 4 will be described in detail.

In step 700, the feature quantity vector is initialized.

In step 702, the processes of steps 704 to 708 are repeatedly executed for all voxels of the image to be processed. That is, as shown in FIG. 5 , all voxels to be processed are sequentially scanned using the lattice structure 506 .

In step 704, the processing from step 706 to step 708 is repeatedly executed for all 251 kinds of dot matrix structures shown in FIG. 6 .

In step 706, it is judged whether all the pixels corresponding to the dot matrix structure to be processed are 1 or not. If the judging result is affirmative, go to step 708 .

In step 708, only 1 is added to the component of the feature vector corresponding to the lattice structure to be processed.

Through the series of processes described above, it is possible to calculate a 251-dimensional feature vector related to the stereo high-order local autocorrelation.

The calculation procedure of the conversion matrix 306 in FIG. 3 will be described below with reference to the flowchart in FIG. 8 .

In step 800 , the processes of steps 400 to 402 are repeatedly executed for one or more images of normal scenes stored in advance in a storage device such as a semiconductor memory for learning.

In step 400 , as shown in FIG. 4 , the motion extraction unit 300 extracts a portion in which motion occurs from the video generated by the video acquisition unit 100 .

In step 402 , as shown in FIG. 4 , the feature amount of the video generated in step 400 is calculated by the feature amount calculation unit 302 .

In step 802, principal component analysis is performed on the calculated set of feature quantities for normal scenes. Principal component analysis is one of the multivariate analysis methods. It is possible to summarize information possessed by multiple variables by generating composite variables called principal components based on several variables in a manner independent of each other. This principal component analysis is a method frequently used in the analysis of multivariate data, and since it is described in detail in, for example, "Simple and Easy-to-understand Multivariate Analysis" (Tokyo Book Publishing), the details are omitted here. illustrate. By performing principal component analysis on a set of 251-dimensional feature quantity vectors, 251 principal components and eigenvalues can be found.

In step 804 , according to the result of principal component analysis in step 802 , the local space with low contribution rate to normal behavior is calculated. The conversion matrix 306 is set as a matrix for converting feature quantity vectors into vectors of the local space.

The calculation process of the local space described in step 804 of FIG. 8 will be described in detail below with reference to FIG. 9 . FIG. 9 shows the cumulative contribution rate of each principal component obtained in the principal component analysis shown in step 802 . The so-called cumulative contribution rate is obtained by adding the contribution rate of each principal component in descending order, and it is an indicator that indicates the amount of information that the previous principal component originally had on the data of the analysis object The size of the extent of the description made. For example, if the cumulative contribution rate 900 up to the third principal component is 90%, it means that the first to third principal components have expressed 90% of the original information content of the data. On the other hand, the remaining 4th principal component to the 251st principal component only account for 10% of the original information content of the data.

It can be seen from the above description that the local space composed of the first principal component to the third principal component has a large contribution rate to the normal behavior. However, the local space composed of the 4th principal component to the 251st principal component has little contribution to the normal behavior.

In this way, by using the cumulative contribution rate as a criterion, a local space with a small contribution rate to normal behavior can be obtained.

Hereinafter, referring to FIG. 10 , the method for evaluating the degree of abnormality described in step 406 of FIG. 4 will be described. The evaluation of the degree of abnormality is carried out in the local space shown in Fig. 9 with a small contribution rate to the normal behavior. This is because, in this local space, the dispersion of the characteristic quantities of normal behavior is small, but the dispersion becomes large in the case of behavior other than normal behavior, that is, abnormal behavior.

Hereinafter, this local space is assumed to be a space composed of principal components below the nth principal component. Originally, the layout space should be a 252-n-dimensional local space, but for the sake of illustration, the principal components with large contribution rates are shown in the form of two axes in Figure 10, namely the nth principal component and the n+1th principal component. A set of feature quantities 1000 is a set of normal data 310 . In the local space where the contribution rate to the normal behavior is small, the feature quantities are distributed around the center of gravity xn1004 of the set as the center. Therefore, if the feature quantity vector x1002 of the image currently being evaluated is near the feature quantity set 1000, it can be judged as normal, and if it is far away, it can be judged as abnormal. Wherein, the distance 1006 between the two is taken as the degree of abnormality.

The distance between the feature quantity vector x1002 and the feature quantity set 1000 can be calculated by using the Euclidean distance calculation method with low calculation cost. However, in this embodiment, the Mahalanobis distance calculation method that takes into account the dispersion of feature quantity sets is used. Assuming that the inverse matrix of the variance covariance matrix of the feature quantity set 1000 is S ^-1 , the Mahalanobis distance can be calculated by the following formula (2).

D ² ＝(xx _n ) ^t S ^-1 (xx _n ) (2)

Hereinafter, referring to the flowchart of FIG. 11 , the processing flow of calculating the judgment threshold 106 by the abnormal behavior detecting device of this embodiment will be described.

In step 1100 , the process of step 1102 is repeatedly executed by the abnormality degree calculation unit 102 for all evaluation scenarios. The so-called evaluation scene is a video library of normal scenes and abnormal scenes, and the result that should be judged by the device is assigned to the evaluation scene.

In step 1102 , the judgment threshold calculation unit 110 calculates the largest abnormality degree of each frame of the evaluation scene to be processed as a representative value of the abnormality degree of the scene.

In step 1104 , the judgment threshold calculation unit 110 generates an error curve to be described later based on the maximum abnormality level of each scene calculated in step 1102 . The so-called error curve is a curve showing how the false positive rate and non-report rate change according to the judgment threshold. Among them, the false positive rate is the ratio of misjudging a normal scene as an abnormal scene, and the non-reporting rate is the misjudgment of an abnormal scene as an abnormal scene. Ratio for normal scenes.

In step 1106 , an appropriate threshold is determined based on the error curve generated in step 1104 . This determination method will be described in detail later.

Referring to FIG. 12 , the calculation sequence of the maximum abnormality degree of each scene described in step 1102 in FIG. 11 will be described in detail. The horizontal axis in the graph represents the frame number, and the vertical axis represents the degree of abnormality. This figure shows how the degree of abnormality of each frame changes when the evaluation scenes are sequentially evaluated. The curve 1200 is the evaluation result of the abnormality degree of each frame of the scene 1 . In the figure, the maximum value of the degree of abnormality is point 1204, and this value is taken as the maximum degree of abnormality of scene 1. Similarly, the curve 1202 is the evaluation result of the abnormality degree of the scene 2, and the abnormality degree corresponding to the point 1206 is taken as the maximum abnormality degree of the scene 2.

The calculation steps of the error curve described in step 1104 of FIG. 11 will be described in detail below with reference to FIGS. 13 and 14 .

FIG. 13 is a bar graph showing the maximum degree of abnormality of each scene calculated in step 1102 of FIG. 11 . In this embodiment, scenes 1 to 7 are normal scenes, and scenes 8 to 14 are abnormal scenes. From this graph, it can be seen that the maximum abnormality level tends to be small for normal scenes and large for abnormal scenes.

Among them, a judgment threshold 1300 is introduced as a criterion for judging whether it is abnormal or normal. When the maximum abnormality degree is less than the judgment threshold 1300, the scene can be regarded as a normal scene, and when the maximum abnormality degree is above the judgment threshold 1300, the scene can be regarded as an abnormal scene.

The anomaly detection performance of the device can be evaluated by false alarm rate and non-alarm rate. The so-called false alarm rate refers to the rate of wrongly judging a normal scene as an abnormal scene, and the smaller the value, the better. In the case of FIG. 13 , among the seven normal scenes, the maximum abnormality degree 1302 of the scene 5 exceeds the judgment threshold 1300 , and it is erroneously judged as abnormal. At this time, the false alarm rate is 1/7=14%. The unreported rate refers to the rate of misjudging an abnormal scene as a normal scene, and the smaller the value, the better. In the case of FIG. 13 , among the seven abnormal scenes, since the maximum abnormal degree 1304 of the scene 12 is smaller than the judgment threshold 1300 , it is wrongly judged as a normal scene. At this time, the unreported rate is 1/7=14%.

The false alarm rate and non-report rate vary with the value of the judgment threshold. Its changes are shown in Figure 14. In the figure, the curve 1400 represents the false alarm rate, and the curve 1402 represents the non-report rate. It can be seen from the figure that there is a trade-off relationship between the false positive rate and the non-reported rate. That is, if a larger judgment threshold is set to reduce the false alarm rate, the non-alarm rate increases. On the contrary, if a smaller judgment threshold is set to reduce the non-report rate, the false positive rate will increase. In step 1106 of FIG. 11 , the judgment threshold 106 is set according to the graph, so as to achieve a more ideal detection performance for the user of the device.

As a candidate for the judgment threshold 106, for example, there is a point 1404 where the false alarm rate and the non-report rate are equal, that is, the equal error rate (Equal Error Rate, EER). At this time, a judgment threshold 1406 corresponding to an equal error rate is set as the judgment threshold 106 .

In addition, the judgment threshold 1408 when the non-report rate is 0 or the judgment threshold 1410 when the false alarm rate is 0 may be set as the judgment threshold.

Furthermore, since the selection conditions of these judgment thresholds differ depending on the purpose and conditions of use of the device, they may be selected by the user of the device.

In addition, the error curve in FIG. 14 may be presented to the user through an output device such as a display. At this time, it may also be set so that the determination threshold can be selected on the screen through an input device such as a mouse or a keyboard.

According to the method described above, there is an effect that the degree of freedom of installation can be increased.

Through the above-mentioned embodiments, it can be automatically determined whether an abnormality occurs according to the degree of abnormality in the image. Therefore, when an abnormality occurs, by issuing an alarm or notifying the security personnel, measures can be taken immediately for the abnormality. Furthermore, it is possible to automatically determine a judgment threshold as a criterion for judging occurrence of an abnormality based on the evaluation result of the scene for evaluation, thereby obtaining an optimum judgment threshold. Therefore, even when the object of abnormality judgment is changed, it is possible to flexibly take corresponding measures only by creating a scene for evaluation.

In the embodiment described above, the local space with a small contribution rate to the normal space shown in Fig. 9 is determined according to the preset cumulative contribution rate, but it is also possible to calculate the local space, the local space can be determined in such a way that the detection accuracy becomes optimal.

The method of determining the local space described above will be described below with reference to FIG. 15 . An EER curve 1500 in FIG. 15 shows the relationship between the EER value and n when calculating the degree of abnormality in the local space constituted by the nth principal component to the 251st principal component. It is generally believed that the smaller the EER illustrated in Fig. 14, the higher the detection performance. Therefore, it is only necessary to select the number n of principal components corresponding to the minimum value 1502 of the EER curve 1500 .

Accordingly, it is possible to automatically determine the local space in such a manner that the detection accuracy is optimized. Therefore, the user can achieve the best detection performance with little effort.

Fig. 16 is an illustration of a monitoring screen using the abnormal behavior detection device of the present invention. This monitoring screen is displayed on a display device of a personal computer or a monitoring terminal with a built-in abnormal behavior detection device. The image in this example is the image captured by the camera inside the elevator car.

Area 1600 is a display area for displaying the target image currently being processed.

A region 1602 is a region for displaying the calculated changes in the degree of abnormality as a trend curve 1604 one by one in a time-series format. The straight line 1606 is the judgment threshold.

Area 1608 is a display area for abnormality judgment results. The area 1608 can distinguish the abnormality judgment result into two states, normal and abnormal, and display which of the two states the abnormality judgment result belongs to according to the content of the judgment result 108 of the current frame.

In addition, instead of being classified into two grades of normal and abnormal, three or more grades may be used to display the degree of abnormality. Area 1608 exemplifies a case in which the degree of abnormality is displayed in three levels. For example, by indicating normality in blue, mild abnormality in yellow, and severe abnormality in red, supervisors can intuitively grasp the situation.

Hereinafter, referring to the flowchart of FIG. 17 , the abnormality judgment processing of multiple levels shown in the area 1608 of FIG. 16 will be described in detail. Here, the ratio of time occupied by the state judged to be abnormal in the period from a certain time point in the past to the current time point is calculated, and a plurality of levels of judgments are performed based on the ratio.

In step 1700, the ratio p of the states judged to be abnormal during the period from a certain time point in the past to the current time point is calculated.

In step 1702, it is compared whether or not the occupancy ratio p is smaller than a predetermined threshold value pb for blue judgment. If the result of the comparison is yes, execute step 1704 . If the result of the comparison is negative, execute step 1706 to step 1710 .

In step 1704, the judgment result is determined to be blue, and the process ends.

In step 1706, it is compared whether or not the occupancy rate p is smaller than a predetermined threshold py for yellow judgment. If the result of the comparison is yes, execute step 1708 . If the comparison result is negative, then execute step 1710 .

In step 1708, the judgment result is determined to be yellow, and the process ends.

In step 1710, the determination result is determined to be red, and the process ends.

In the above-mentioned processing, the case where the judgment result is displayed in three levels has been described, but it is also possible to display the judgment result using more levels. At this time, it is only necessary to pre-set the thresholds of each level. In this way, there is an effect that a convenient monitoring method can be provided for the monitor by displaying the judgment result in a plurality of levels.

According to the above-mentioned embodiments, it is possible to conveniently grasp the correspondence between the image content of the evaluation object and the abnormality degree of the image. Therefore, for example, even if there is a false alarm, the monitor can easily and quickly confirm whether an abnormal behavior has really occurred through the video.

Fig. 18 shows an elevator apparatus provided with an abnormal behavior detection device according to the present invention. However, in FIG. 18 , illustration of drive mechanisms such as hoists, slings, and counterweights is omitted. The image inside the elevator car is acquired by the camera 30 arranged in the elevator car 40 , and the image signal is transmitted to the abnormal behavior detection device 10 arranged in or outside the elevator passage through the tail cable 50 . When an abnormality is detected in the image in the elevator car 40, the abnormal behavior detection device 10 outputs a judgment result signal indicating the abnormality to the elevator control device 20 installed in the hoistway or the machine room. After the elevator control device 20 receives the judgment result signal indicating abnormality, it controls the hoisting machine so that the elevator car 40 stops on the nearest floor, and controls the elevator door driving device to open the elevator car 40 and the doors of the elevator hall. Alternatively, the control device 20 activates an alarm device such as an alarm installed in the elevator car 40 . According to the above-mentioned elevator equipment, it is possible to quickly leave people who are in the same elevator car away from those who behave abnormally, or to suppress abnormal behavior in the elevator car, thereby improving the safety of the elevator equipment. In addition, the abnormal behavior detection device 10 may also be installed in the elevator car 40 . At this time, the judgment result signal is transmitted to the control device 20 via the tail cable 50 .

Claims (9)

1. abnormal behaviour sniffer has:

The image acquisition unit, it obtains the image of monitored object;

The intensity of anomaly calculating part, its intensity of anomaly to the image that described image acquisition unit is obtained calculates; And

Unusual judging part, it is according to threshold value, and the intensity of anomaly that calculates from described intensity of anomaly calculating part judges whether to have taken place abnormal behaviour,

In the described abnormal behaviour sniffer, also have make the expression rate of false alarm and not the error curve of the relation between newspaper rate and the described threshold value be presented at display part on the picture.

2. abnormal behaviour sniffer as claimed in claim 1 is characterized in that,

Also has the input mechanism of on the picture that shows described error curve, selecting described threshold value.

3. abnormal behaviour sniffer as claimed in claim 1 is characterized in that,

Described intensity of anomaly calculating part has:

The activity extraction unit, it extracts the part that has produced motion from described image;

Feature value calculation unit, it calculates first characteristic quantity of the image that is generated by described activity extraction unit;

The characteristic quantity converter section, it is converted to second characteristic quantity in the linear transformation mode with described first characteristic quantity; And

Intensity of anomaly Rating and Valuation Department, it compares described second characteristic quantity and pairing the 3rd characteristic quantity of normal behaviour and calculates intensity of anomaly.

4. abnormal behaviour sniffer as claimed in claim 3 is characterized in that,

Described feature value calculation unit adopts the local autocorrelation characteristic of three-dimensional high order to calculate described first characteristic quantity.

5. abnormal behaviour sniffer as claimed in claim 3 is characterized in that,

Intensity of anomaly calculates according to the Euclidean distance between the center of gravity of the set of described second characteristic quantity and described the 3rd characteristic quantity in described intensity of anomaly Rating and Valuation Department.

6. abnormal behaviour sniffer as claimed in claim 3 is characterized in that,

Intensity of anomaly calculates according to the mahalanobis distance of the set of described second characteristic quantity and described the 3rd characteristic quantity in described intensity of anomaly Rating and Valuation Department.

7. as any described abnormal behaviour sniffer of claim 1 to 6, it is characterized in that,

Have the judgment threshold calculating part, this judgment threshold calculating part is set described unusual judging part required described threshold value when carrying out judgment processing in the mode that can obtain best judged result.

8. abnormal behaviour sniffer as claimed in claim 7 is characterized in that,

The setting of described judgment threshold calculating part and rate of false alarm and any one corresponding threshold in the equal error rate that equates of newspaper rate, 0% rate of false alarm and 0% the not newspaper rate not.

9. as any described abnormal behaviour sniffer of claim 3 to 6, it is characterized in that,

Described characteristic quantity converter section, with rate of false alarm and the mode of the equal error rate minimum that equates of newspaper rate generate the transition matrix of described linear conversion.

Images